Skin Cancer Classification (CNN Baselines)
CNN-based framework for skin lesion and melanoma classification using dermoscopic image datasets, serving as a strong baseline for medical imaging research.
This project explores skin cancer classification using convolutional neural network (CNN) baselines applied to dermoscopic images. The focus is on building a clean, reproducible training and evaluation pipeline that serves as a foundation for more advanced architectures and ensemble methods in medical imaging. This study establishes a strong baseline for dermoscopic image classification, enabling systematic comparison with advanced deep learning models in subsequent research.
The implementation emphasizes correct data handling, model training stability, and metric-driven evaluation, which are key requirements for clinically oriented machine learning workflows. While intentionally kept as a baseline study, the project is designed to scale toward deeper networks, explainability, and external validation. This baseline framework provides a reproducible foundation for benchmarking advanced architectures in medical image classification.
Pipeline Overview
End-to-End CNN Workflow
- Data Preparation: loading, resizing, normalization, and dataset splitting.
- Model Training: CNN baseline architectures trained with supervised learning.
- Evaluation: performance assessment on held-out test data.
Key Components
- Custom data loaders for dermoscopic images
- CNN baseline architectures implemented in PyTorch
- Training and validation loops with logging
- Standard classification metrics and visual diagnostics
- Extensible structure for future model upgrades
Evaluation & Metrics
- Accuracy for overall predictive performance
- Per-class Precision, Recall, and F1 to capture class imbalance
- Confusion Matrix for diagnostic error analysis
- Visual inspection of model predictions on representative samples
Results & Analysis
The CNN baselines provide a strong performance baseline for dermoscopic image classification. Results highlight both the strengths and limitations of shallow architectures, motivating the use of deeper networks, transfer learning, and ensemble strategies in subsequent work.
Sample Lesion Classification
This baseline study forms the groundwork for advanced medical imaging experiments, including transfer learning, explainability (Grad-CAM), and multi-model ensembles.
Repository
The full codebase, training notebooks, and evaluation scripts are available at:
https://github.com/md-naim-hassan-saykat/skin-cancer-cnn